AMPA-Net

This is the official TensorFlow implementation of AMPA-Net [Nanyu Li, Charles C. Zhou, AMPA-Net： Optimization-Inspired Attention Neural Network for Deep Compressed Sensing， In the IEEE 20th International Conference on Communication Technology (ICCT) oral] which can be downloaded from https://arxiv.org/abs/2010.06907

Abstract

Compressed sensing (CS) is a challenging problem in image processing due to reconstructing an almost complete image from a limited measurement. To achieve fast and accurate CS reconstruction, we synthesize the advantages of two well-known methods (neural network and optimization algorithm) to propose a novel optimization-inspired neural network which dubbed AMP-Net. AMP-Net realizes the fusion of the Approximate Message Passing (AMP) algorithm and neural network. All of its parameters are learned automatically. Furthermore, we propose an AMPA-Net which uses three attention networks to improve the representation ability of AMP-Net. Finally, We demonstrate the effectiveness of AMP-Net and AMPA-Net on four standard CS reconstruction benchmark data sets.

Compatibility

• The code is tested using Tensorflow 1.0 under Ubuntu 16.04 with Python 2.7.

• Recommend Environment: Anaconda

Preparing Training Datasets-Image-91

Download Training_Data_Img91.mat from：https://pan.baidu.com/s/1UgRuDbIXCNZOEuedVlK8bA?pwd=cq1b code is cq1b and cp it into this folder

Preparing Testing Datasets:Urban-100 and BSDS100

Download Urban-100 from：https://pan.baidu.com/s/1AUNWfVx8Jy12D12yJClWYg?pwd=rr3q code is rr3q and cp it into this folder

Download BSDS-100 from：https://pan.baidu.com/s/1KwpRYlp_7SjphA_Vc2LMiA?pwd=kbyw code is kbyw and cp it into this folder

Running training and testing

Python trainn.py

Results

Our model achieves the following performance on Set11, BSD68, BSDS100, Urban100

performance on Set11

Model name	cs=50%	cs=40%	cs=25%	cs=10%	cs=4%	cs=1%
BM3D-AMP	34.11	32.79	27.87	22.12	17.32	4.91
LD-AMP	35.92	33.56	28.46	22.64	18.40	5.21
Ad-Recon-Net	34.21	32.72	30.80	27.53	23.22	20.33
FGMN					23.87	21.27
Full-Conv			32.69	28.30		21.27
DR2-Net	32.40	31.20	28.66	24.71	20.08	17.44
ISTA-Net	37.43	35.36	31.53	25.80	21.23	17.30
ISTA-NetPlus	38.07	36.06	32.57	26.64	21.31	17.34
AMP-Net	39.52	37.13	33.60	28.47	24.21	20.48
AMPA-Net	40.32	38.27	34.61	29.30	24.95	21.59

performance on Set68

Model name	cs=50%	cs=40%	cs=25%	cs=10%	cs=4%	cs=1%
LD-AMP	31.22	30.12	22.79	20.11	12.02	3.50
Ad-Recon-Net	29.94	28.82	25.02	25.45	22.28	19.68
ISTA-NetPlus	34.01	32.21	29.21	25.33	22.17	19.50
AMP-Net	35.02	33.10	30.25	26.92	23.77	20.85
AMPA-Net	36.33	34.41	31.38	27.58	24.90	21.99

performance on BSDS100

Model name	cs=50%	cs=40%	cs=25%	cs=10%	cs=4%	cs=1%
LD-AMP	30.81	29.57	22.46	19.64	10.40	3.21
Ad-Recon-Net	29.21	28.12	24.80	25.13	22.22	19.35
ISTA-NetPlus	33.64	31.83	29.00	25.08	22.10	19.17
AMP-Net	34.21	32.13	29.59	26.87	23.21	19.48
AMPA-Net	35.95	34.03	31.01	27.29	24.75	21.62

performance on Urban-100

Model name	cs=50%	cs=40%	cs=25%	cs=10%	cs=4%	cs=1%
LD-AMP	30.41	29.12	22.02	17.14	8.42	1.31
Ad-Recon-Net	29.15	27.90	24.20	23.13	19.22	16.82
ISTA-NetPlus	33.94	31.96	28.32	23.44	19.41	16.47
AMP-Net	34.08	31.95	29.02	26.20	20.01	16.88
AMPA-Net	35.86	33.92	30.49	25.76	22.40	18.86

Contact

Email: [email protected]

License

MIT